Every job I have had in energy has involved the same painful week. A new rule drops, and someone has to read it, work out what it actually asks for, and write that down in a way a team can build. Usually that someone was me, and usually it took far longer than it should have.
So I built a tool to do the first read. It is called RegOps, and it runs inside MyOpsLab. You give it a regulation as a PDF, and it hands back first drafts of the documents I would normally write by hand.
Why does turning a new energy rule into requirements take so long?
Rules do not arrive as requirements. A standard AEMC rule change takes about six months from the consultation paper to the final rule, and what comes out the other end is a legal document written for the whole market, not for your product.
Take flexible trading. The AEMC made the final rule on 15 August 2024. The parts for in-built metering on things like streetlights and public EV chargers had to be in place by 31 May 2026, and most of the rest by 1 November 2026. That gave the industry just over two years. It sounded like plenty. It was not, once you counted the reading, the impact assessment, the build and the testing, across every team that touched billing, metering and market systems.
Or take the smart meter rule. The AEMC made it final on 28 November 2024, with the aim of a smart meter for every customer by 2030. It sounds like one change. It was really two rule books changed at once, the National Electricity Rules and the National Energy Retail Rules, with parts switching on at five different dates between 5 December 2024 and 1 July 2026. So the first read was never just 'what do we have to do'. It was 'which of these is the retailer's job, which belongs to the metering coordinator or the network, and which date does each one start'. Get one start date wrong and you are either late or building something more than a year too early.
The slow part is the first pass. Finding every "must", working out who it applies to and when, and mapping it to what the product does today.
What should an obligations register capture?
An obligation is one thing someone must do, by a certain time, for a certain customer. Each row in the register needs enough to act on and enough to prove later:
- the clause it comes from, word for word
- the obligation in plain words
- who it applies to, such as the retailer, the network or the metering coordinator
- when it starts and what triggers it
- the retailer requirement, meaning what the retailer has to do about it
- the platform requirement, meaning what the software has to do so the retailer can
- how you will show you did it
The last one, how you will show you did it, is the one people skip. It is also the one an auditor asks about first.
Splitting the retailer requirement from the platform requirement matters. The rule puts the duty on the retailer, but the retailer can only meet it if the platform does its part.
What does the tool produce?
From one regulation, it drafts:
- an obligations register, so nothing in the rule slips through
- a traceability matrix that links each obligation to where it is handled
- a gap analysis, user stories and a test plan
I have used it on more than 60 regulations so far. Together they come to more than 15,000 pages.
End to end, getting a new rule from the page to a team that can build it used to take weeks. My own first read and draft of a typical rule change went from about 3 days to about 3 hours.
Where does AI get it wrong?
It writes a first draft, not the final word. It is good at finding obligations and not so good at judgement.
Here is a real one. On one rule change, the wording of the final rule was more relaxed than what the AEMC said it intended in the final determination. The tool read the rule, did what the rule said and missed the gap between the two. I missed it too. A retailer reviewing my document caught it. The rule tells you what you must do. The determination tells you what the regulator meant. You need to read both, and so does the tool.
Because every obligation links back to its clause, checking the draft is quick. You read the clause, you read the row, and you decide if they match.
I check every high risk obligation against its clause myself, things like life support, billing and customer notices, and sample the rest. Then the draft goes to the people who live with the rule every day. They catch what a tool and one reader will not.
Where did the idea go next?
At Kaluza I took the same idea further, on Kaluza's own systems, and spent most of my effort on making the AI's output trustworthy. I wrote about that in making an AI assistant tell the truth.
What still needs a person?
It does not know which obligations carry real risk, which ones the platform already handles, or where the business wants to draw the line. That part is still mine, and I think it should stay that way.
The point was never to remove the person. It was to clear the boring part out of the way so the person can spend their time on the decisions that actually matter.